Adverse Drug Reaction Predictions Using Stacking Deep Heterogeneous Information Network Embedding Approach.

Area of Science:

• Computational pharmacology
• Bioinformatics
• Drug discovery

Background:

• Predicting adverse drug reactions (ADRs) is crucial for drug safety and development.
• Existing computational methods often overlook the impact of protein-protein interactions (PPIs) on drug targets.
• Current approaches are frequently task-specific, limiting their generalizability.

Purpose of the Study:

• To develop a novel, generic approach for learning drug representations that incorporates PPI information.
• To enhance the prediction of adverse drug reactions (ADRs) by leveraging heterogeneous drug data.
• To create a model applicable to diverse ADR prediction tasks.

Main Methods:

• Proposed a heterogeneous network embedding approach named SDHINE.
• Designed meta-path-based proximities, including a target propagation proximity utilizing PPI networks.
• Constructed a semi-supervised stacking deep neural network optimized by meta-path proximities.

Main Results:

• SDHINE demonstrated superior performance compared to state-of-the-art network embedding methods on three ADR prediction tasks.
• Experiments confirmed the effectiveness of integrating PPI information into drug embeddings.
• Visualization of drug representations showed enhanced drug differentiation.

Conclusions:

• SDHINE offers a powerful and versatile framework for ADR prediction by effectively integrating PPI data.
• The proposed method advances computational pharmacology by improving drug representation learning.
• This approach holds significant potential for improving drug safety and discovery pipelines.